EXAONE 3.5 Released as Open Source with 3 Models - Achieves Frontier AI-level Performance, Top-Tier Instruction Following and Long Context

Following the release of EXAONE 3.0 last August, LG AI Research has released 3 base models based on EXAONE 3.5 as open source, reflecting feedback demanding a model lineup tailored to diverse use cases.

The released models consist of three types as follows.

• 2.4B model: An ultra-lightweight model that can run even in on-device and low-spec GPU environments
• 7.8B model: A lightweight model for general-purpose use, with improved performance compared to the previous version
• 32B model: A Frontier AI-level high-performance model for users who prioritize performance above all

EXAONE 3.5 has significantly improved Long Context understanding and processing capability. All models support an Effective Context Length of 32K tokens, and are highly effective when using RAG (Retrieval-Augmented Generation) technology. In particular, it has demonstrated top-tier performance in both Korean and English.

Additionally, Instruction Following capability has been strengthened to enhance usability in actual industrial settings. All three models ranked first in average score across 7 benchmarks, significantly outperforming global models of the same size. It also shows SOTA (State-of-the-Art)-level performance in general domains such as math and programming ability.

LG AI Research is currently discussing the application of AI solutions through partnerships with Polaris Office, Hancom, and others, and plans to make business outcomes visible through this and contribute to the development of the AI ecosystem.